Skip to content

Open source Grok Bot alternative, Bring your own agent and subscription: message Claude Code, Codex, Gemini CLI, Cursor, Copilot and 40+ AI agents as bots on your own computer, from iPhone, Mac, Windows, Linux or terminal. Agent-agnostic and free. Also a Muse and Dots alternative. Built in Rust.

3.7 KB
Raw Blame

Voice call transcripts

Status: superseded and shipped as stage 1 of voice calls owned by the host (2.25.0): transcripts, the titled call line, hidden call messages and dreaming over calls, with the host running the call. Current behavior: voice call. The original plan stays below for the record.

Goal

  1. Every call keeps its transcript: what the user said and what the voice answered, in order.
  2. The call's line in the chat becomes a short title that sums the call up ("Plans the 2.26 release · 03:12"), and opens the transcript.
  3. Dreaming reads the transcript like the rest of the chat, so what the user said on a call reaches long-term memory.

On-device and speech-to-text calls already send each utterance as an ordinary message; they get the title (2) and nothing else changes for them.

Steps

1. Capture (iPhone, Android, desktop)

  • Gemini Live: inputAudioTranscription and outputAudioTranscription in the setup; the server sends serverContent.inputTranscription / outputTranscription pieces, joined per turn (checked on a phone on 2026-10-25: accurate, but Chinese comes back in Simplified script for a Traditional speaker).
  • OpenAI Realtime: input transcription in session.update; conversation.item.input_audio_transcription.completed and response.output_audio_transcript.done.
  • Each engine keeps [{from: "user" | "voice", text, at}] for the call. Chinese goes through the reader's script (VoiceSettings.inReadersScript on iPhone, its ports elsewhere). A turn cut short by barge-in keeps what was said.

2. Host: the call entry

  • logCall takes an optional transcript (additive; older apps send none). The host keeps it on the call's notice entry (data.transcript) and validates it (length caps, from limited to the two values).
  • A one-shot agent of the bot's harness (chat::oneshot, as chat::naming uses) writes a title of a few words in the user's language from the transcript; the host puts it in the entry's text (it stays "Voice chat" until then, and when there's nothing to sum up) and emits the update. Older apps already show text · mm:ss, so they get the title for free; no minApp change.

3. Memory

  • dream::message expands a call entry: the user's lines count as the user's own words (spoken, not pasted), the voice's lines are context only. A call counts its user lines toward the eight-message trigger.
  • send_to_bot messages are written by the voice model, not typed by the user (it may translate or rephrase), so the apps mark them (from: "voice") and dreaming quotes the transcript instead.
  • Update the "Voice transcripts aren't marked yet" limit in context and memory.

4. The chat line (iPhone, Android, desktop, terminal)

  • The line shows title · 00:26; tapping it opens the transcript (a sheet on iPhone and Android, the desktop Sheet, a key in the terminal). Without a transcript it isn't tappable.
  • States: "Voice chat" while the title is being written; the transcript sheet lists "You" and the voice's lines.

5. Ship

  • Minor version (protocol addition). WhatsNew for iPhone and Android, the voice-call and memory docs, the user manual's voice page. Tests: transcript validation and title on the host, dreaming over a call entry, each client's transcript parsing; e2e a Gemini call on a phone and on the desktop.

Open questions

  • The transcript under the call line (planned) or each utterance as its own chat bubble?
  • The title from the bot's harness (planned: same as naming, works for on-device calls) or from the voice provider's key?
  • Thumbs up / down on the call line, for rating the voice?